smithy-agent

Unattended worker that runs on a Windows machine, registers itself with a smithy-cloud orchestrator, polls it for commands, and executes deployed process bundles (Python code) in isolated per-process venvs.

One-line install (Windows)

Regular PowerShell (it self-elevates and installs Python 3.12 via winget if missing):

irm https://raw.githubusercontent.com/as-kurosss/smithy-agent/master/install-agent.ps1 | iex

With parameters:

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/as-kurosss/smithy-agent/master/install-agent.ps1))) `
    -Orchestrator https://cloud.example.com -Name prod-1 -JoinToken <TOKEN>

Manual install

pip install smithy-agent

Quick start (manual, foreground)

pip install smithy-agent
smithy-agent --orchestrator http://orch-host:8000 --name my-agent \
    --url http://this-machine:8001

--url must be the address at which the orchestrator can reach this machine (the orchestrator pushes commands and files to it).

Unattended installation (recommended)

UI automation needs an interactive desktop, so the agent runs as a scheduled task inside a user's log-on session — not as a session-0 service. Use a dedicated local user with auto-logon for production agents.

# one command: writes config + creates the scheduled task
smithy-agent-service install \
    --orchestrator http://orch-host:8000 \
    --name prod-1 \
    --join-token <AGENT_JOIN_TOKEN> \
    --user "WORKGROUP\\rpa-bot"     # optional, default: current user

smithy-agent-service start        # launch now
smithy-agent-service status       # task state
smithy-agent-service stop | uninstall

What you get:

• task trigger at log on of the automation user → agent starts with the desktop, no manual steps after reboots;
• Task Scheduler restart policy (up to 999 retries, 1-minute interval) and, belt and braces, a crash-restart loop inside the agent with exponential backoff (5 s → 60 s, reset after a stable 5-minute run);
• config at %LOCALAPPDATA%\smithy_agent\config.json, rolling log at %LOCALAPPDATA%\smithy_agent\agent.log.

Registration honours AGENT_JOIN_TOKEN (pre-shared in the orchestrator settings) or an admin JWT — whichever you pass with --join-token.

Privacy note: failure screenshots

On a failed run the agent (with the [screenshot] extra installed) captures the entire virtual screen — all monitors and every visible window, not just the automated app — and uploads it to the orchestrator. Keep this in mind on shared/multi-user machines: anything on screen at the moment of failure (other apps, browser sessions, passwords) ends up in the cloud.

Deploy with the plain smithy-agent package if screenshots are unwanted; the agent runs fine without the extra and simply skips capture.